A large model hallucination-free output verification and integrated AI dual-mode inference system and method

Abstract

This invention discloses a large-scale model-based, hallucination-free output verification and integrated AI dual-mode inference system and method, belonging to the field of artificial intelligence and intelligent interaction technology. This invention deeply integrates deterministic precise inference and large-scale model open generation into a single system, supporting users to choose between precise and open response modes. Precise responses are based on standardized authoritative data for deterministic comparison, outputting verifiable content; open responses are generated by the large model and undergo real-time verification to eliminate error risks. The system supports direct output before processing, post-processing verification, and dual-mode switching, and can be deployed in multiple scenarios, achieving low-cost, reliable, and universally applicable large-scale models. Its innovative approach makes it suitable for high-reliability AI applications such as government affairs, finance, and healthcare.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, large model security, trusted interaction, and rational intelligent agents. Specifically, it relates to an AI system and method that deeply integrates deterministic authoritative reasoning with large model generation and supports users to choose their own answering methods. Background Technology

[0002] Existing large-scale models, retrieval enhancement, and content verification technologies have the following inherent drawbacks: • Large models are generated based on probability, which inevitably leads to illusions, fabrications, and misinformation, making them unsuitable for high-confidence scenarios; • Pure search systems only output fixed content and lack the ability to create, chat, or express themselves openly; • Verification is mostly done externally, without deep system-level integration; • Users cannot choose between "precise and reliable" or "open and flexible"; • Open generation lacks security safeguards and is prone to errors, misleading information, and illegal content; • Large-scale APIs are billed based on the number of calls, which is costly and cannot be deployed privately.

[0003] This invention fundamentally solves the above problems through dual-engine system-level integration, user-autonomous mode switching, and deterministic content verification as a fallback, and is essentially different from existing technical approaches. Summary of the Invention

[0004] 1. System Overall Architecture The system includes two deeply integrated engines: • Deterministic and precise reasoning engine: Based on pre-built, verified, and traceable standardized authoritative data, it outputs authentic and reliable content through feature extraction, matching, and comparison, without probability, generation, guessing, or fabrication.

[0005] • Large-scale open generation engine: used for text generation, creation, expansion, and chat. Content security and authenticity must be verified before output.

[0006] 2. User-selectable dual-mode Users can switch freely within the same interface and context: • Precise Answer Mode: Striving for 100% authenticity, verifiability, and no fabrication; • Open-ended response mode: encourages creativity, expression, and flexible expansion, but must undergo security verification.

[0007] 3. Precise Answer Mode step: 1. Receiving user issues; 2. Extract keywords, entities, conditions, and constraints from the problem; 3. Perform deterministic matching and comparison within standardized authoritative data; 4. Match successful → Output true, verifiable, and standardized content; 5. Matching failure → Clearly state "no reliable information", do not guess or fabricate.

[0008] Features: Does not rely on large model inference, zero calling cost, and 100% verifiable content.

[0009] 4. Open-ended response mode step: 1. Users select the open mode; 2. Large model execution, generation, creation, and expansion; 3. The system performs feature decomposition and authenticity comparison on the output content; 4. Detect errors / false information / contradictions → automatically intercept and correct; 5. Verification successful → Output final content.

[0010] Characteristics: Flexible expression, but never spreads misinformation.

[0011] 5. Two core working mechanisms • (1) Pre-processing direct output mechanism (precise mode): User question → feature extraction → authoritative data matching → output standard answer; large models can be selected to only perform formatting, without generating new information, zero illusion, and zero cost.

[0012] • (2) Post-verification mechanism (open mode fallback): Large model output → content decomposition → authenticity comparison → error interception → correction → secure output.

[0013] 6. Deployment Mode • Cloud deployment • Local private deployment • Offline deployment without internet access • Embedded device deployment 7. Core Technology Effectiveness • Precise mode: 100% authentic, verifiable, no fabrication, zero cost of invocation; • Open mode: Flexible creation, with a safety net, and no spread of misinformation; • User-defined choices, unified experience; • Compatible with all large models, requiring no training or modification; • Secure, compliant, and reliable across all scenarios. Detailed Implementation

[0014] Example 1: Precise Mode A user asked: "How long does the high-speed rail journey from Beijing to Shanghai take?" The system extracts keywords, matches them with standardized authoritative data, and outputs verifiable answers.

[0015] Example 2: Open Mode The user said, "Help me write a travel description for a trip from Beijing to Shanghai." Large model generation → System verification of content accuracy → Output.

[0016] Example 3: Illusion Interception The large model outputs an error message → the system fails the comparison → it is intercepted and replaced with real content.

Claims

1. A large model hallucination-free output verification and integration type AI dual mode inference system and method, characterized in that, The system deeply integrates a deterministic and precise reasoning engine with a large-scale open generation engine, allowing users to choose between precise or open response modes. The system enables front-end direct output, post-end verification, and user-controllable dual-mode switching, ensuring a process free of illusions, fabrications, and false information.

2. The system and method of claim 1, wherein, The precise answer mode includes: receiving user questions → extracting features → performing deterministic matching and comparison in pre-built standardized authoritative data → outputting true, verifiable, and unfabricated content; the precise mode does not rely on large model inference, achieving zero calling cost of large models.

3. The system and method of claim 1, wherein, The open response mode includes: text generation, creation, or expansion of expression performed by a large model; and content authenticity and security verification before output, blocking illusory, erroneous, and illegal content.

4. The system and method according to claim 1, characterized in that, Users can manually switch between precise answer mode and open answer mode with a single click on the same interface and within the same context.

5. The system and method according to claim 1, characterized in that, The aforementioned pre-processing mechanism includes: user question → feature extraction → authoritative data matching → output of standard answer; the answer can be optionally sent to a large model for formatting, and the large model does not participate in factual reasoning.

6. The system and method according to claim 1, characterized in that, The post-verification mechanism includes: receiving the output content of the large model → feature decomposition → comparison with standardized authoritative data → determining errors → intercepting and correcting → outputting a security result.

7. The system and method according to claim 1, characterized in that, System deployment methods include: cloud, local private deployment, offline deployment, and embedded devices.

8. The system and method according to claim 1, characterized in that, The output of the accurate answer mode is verifiable and traceable, ensuring its authenticity.

9. The system and method according to claim 1, characterized in that, The system is compatible with all generative large models, requiring no model modification, training, or fine-tuning.

10. The system and method according to any one of claims 1-9, characterized in that, It is applied to government affairs, finance, law, healthcare, education, enterprise services, rational intelligent agents, and personal intelligent assistants.